As more devices become interconnected and users consume more data, the demand placed on servers accessed by users has grown commensurately and shows no signs of letting up in the near future. Among others, these demands include increased data transfer rates, switching architectures that require longer interconnects, and extremely cost and power competitive solutions.
There are many interconnects within server and high performance computing (HPC) architectures today. These interconnects include within blade interconnects, within rack interconnects, and rack-to-rack or rack-to-switch interconnects. In today's architectures, short interconnects (for example, within rack interconnects and some rack-to-rack) interconnects are achieved with electrical cables—such as Ethernet cables, co-axial cables, or twin-axial cables, depending on the required data rate. For longer distances, optical solutions are employed due to the very long reach and high bandwidth enabled by fiber optic solutions. As new architectures emerge, such as 100 Gigabit Ethernet, traditional electrical connections, however, are becoming increasingly expensive and power hungry to support the required data rates. For example, to extend the reach of a cable or the given bandwidth on a cable, higher quality cables may need to be used or advanced equalization, modulation, and/or data correction techniques employed which add power and latency to the system. For some distances and data rates required in proposed architectures, there is no viable electrical solution today. Optical transmission over fiber is capable of supporting the required data rates and distances, but at a severe power and cost penalty, especially for short to medium distances, such as a few meters.